EXPLORING THE ASSOCIATION BETWEEN ETHIOPIAN KIGNIT SCALE AND YARED ZEMA MODE CLASSIFICATIONS USING ARTIFICIAL INTELLIGENCE

Abstract

Ethiopian musicologists are exploring the relationships among four pentatonic scales: Tizita, Bati, Anchihoye Lene, and Ambasal, alongside the three scales of Yared's Zema from the EOTC: Geez, Ezl, and Araray. However, the traditional methods of studying the association have been limited in their ability to capture complex patterns and variations. Therefore, there is a need to leverage AI techniques to address this gap and gain a deeper understanding of the relationship between those two scales. A comprehensive review of prior studies clearly shows that there is a lack of well prepared datasets for machine learning, cross-validating unsupervised with supervised learning indicating there is limited intervention of AI technology in the Ethiopian musical association. As a result, experts were unable to obtain the assistance of the AI technology on analysis for scaling systems, preventing the development of the country's music culture. This study aimed to examine the relationship between Ethiopian Kignit and Yared Zema Mode using AI techniques. The study utilized 7,000 audio segments collected from the EOTC's spiritual music and instrumental schools, creating two balanced datasets: SL-EPKMIR and SL-YZMMIR. The study also addressed multi label classification for Ethiopian Kignit with the ML-RPKMIR dataset utilizing chromatic note intervals. This research followed a convergence of exploratory and constructive research design and used python as a programming language for feature extraction and designing a model, sklearn and keras for modeling and mxied approach for data collection. The researchers applied K-means clustering with k values of 3 and 4 to analyze the data and selected features using filter, wrapper, and embedded techniques for classification models. Utilizing a purposeful split of the dataset, the Yeneta model, which employs LightGBM, achieved 83% accuracy for SL-YZMMIR, while the Zemariw model, based on CatBoost, reached 85.7% accuracy for SL-EPKMIR. Although clustering proved challenging, classification results indicated new correlations: Geez with Bati, Ezl with Anchihoye, and Araray with Tizita, while Ambasal remained ambiguous. The GRU model excelled in multi-label classification, achieving a Jaccard score of 94%. Additionally, the researchers developed a new knowledge-based Kignit class for ML-EPKMIR and created an Android application to incorporate this Kignit class.